Early Prediction of Diabetic Macular Edema via Machine Learning Survival Analysis on Checkup Data.

PURPOSE: To predict the risk of diabetic macular edema (DME) onset and to identify features of the risk subgroups. DESIGN: Population-based observational study with a case-control design. SUBJECTS AND CONTROLS: Health checkup and diagnosis data from the JMDC claims database (January 2005-July 2020), one of the largest Japanese epidemiological databases, were used. From 272 337 individuals diagnosed with type 2 diabetes (International Classification of Diseases, 10th Revision E11), we analyzed 2368 pairs of DME and non-DME individuals, which matched 1:1 by the balancing score calculated from regression analysis of DME with sex, age, months of observation, number of checkups, duration of diabetes, and months until the first checkup. METHODS: We employed a multivariate Cox proportional hazards model, regularized Cox models, and a random survival forest (RSF). These models were trained via ID-level bootstrap resampling using 43 health checkup variables (missing ratio <50%) and 404 high-incidence diseases within 6 months, along with age, sex, and duration of diabetes mellitus, to assess DME risk. Temporal changes in RSF-predicted risk scores were analyzed using nonlinear modeling techniques. MAIN OUTCOME MEASURES: Concordance index (C-index), integrated Brier score (IBS), and cumulative/dynamic mean area under the receiver operating characteristic curve (AUC). RESULTS: Thirteen checkup items and 44 disease history variables were significantly associated with the onset of DME. The RSF identified 43.8% of DME cases >5 years prior to onset, with a specificity of 85.5%. The RSF achieved median C-index, IBS, and mean AUC values of 0.694 (95% confidence interval, 0.688-0.697), 0.181 (0.179-0.184), and 0.750 (0.739-0.756), respectively, outperforming both multivariate and univariate Cox models. Three distinct DME risk subgroups were suggested by temporal changes in the RSF-predicted risk score. Predictors of DME onset varied markedly among these subgroups. In the explicit high-risk subgroup, urinary protein and urinary sugar were highly important, and liver function-related blood tests, such as alanine transaminase and γ-gultamyltransferase, were also ranked high in variable importance metrics. Anemia-related laboratory tests were associated with DME development only in this subgroup. CONCLUSIONS: Random survival forest demonstrated superior performance in relative risk prediction of DME using health checkup data. External validation remains an essential prerequisite before any clinical application. FINANCIAL DISCLOSURES: Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.